Help:Ephemera Agent/Configuration

API Keys File

Rather than pasting API keys every session, you can create a local ephemera-keys.json file and load it with one click per session. Keys are read into browser memory only — they are never sent to the server or written to disk.

File Format

<syntaxhighlight lang="json"> {

 "providers": {
   "anthropic": "sk-ant-YOUR-KEY-HERE",
   "openai":    "sk-YOUR-KEY-HERE",
   "gemini":    "AIzaYOUR-KEY-HERE"
 },
 "custom_endpoints": [
   {
     "name":    "Groq Llama",
     "url":     "https://api.groq.com/openai/v1/chat/completions",
     "model":   "llama-3.3-70b-versatile",
     "api_key": "gsk_YOUR-KEY-HERE"
   }
 ]

} </syntaxhighlight>

Delete entries for providers you don't use. The custom_endpoints array is optional.

Loading the File

1. In the LLM tab, click 📂 Choose keys file
2. Select your ephemera-keys.json
3. The LOADED KEYS section shows which providers were found (first 8 characters of each key for verification)
4. The API key field for the current provider auto-fills
5. Switching providers auto-fills that provider's key

Settings Tab

LLM Tiers

Two independent model pickers: Planner and Generator. The planner handles task classification and entity extraction — use a fast, cheap model. The generator handles actual content creation — use the best model available.

Dropdowns show only providers that currently have a key loaded. Settings persist via localStorage.

Red Link AutoGen

Enable red link detection

When enabled, red links in the browser preview pane are highlighted with a dashed underline and a ⚡ indicator.

Click to generate (explore mode)

Clicking a red link triggers the agent to create that page immediately. The new page loads in the preview pane. The wiki grows as you browse.

Batch only

Red links are highlighted but clicking does nothing. A ⚡ Generate N red links button appears in the browser toolbar to process them all sequentially.

Context depth

Controls how much wiki context is gathered before generation. Minimal uses only the World Bible and page title (works immediately). Standard and Deep require the context assembly PHP pipeline to be deployed.

Page type to generate

Auto-detect, Encyclopedia article, or Source document.

Add real-world Wikipedia context

Fetches the introductory paragraph of the matching Wikipedia article for each named entity. Labelled as out-of-universe reference — for scientific and historical grounding only. Off by default.

Context Assembly

Placeholder controls for context assembly settings (active once agent.php is deployed). Currently these values are controlled via Instructions:Config on the wiki.

Wiki-Native Configuration Model

Ephemera Agent follows a wiki-native configuration model: important runtime settings are stored on wiki pages where they can be reviewed and edited without touching PHP files.

Instructions:Config

The Instructions:Config wiki page controls all runtime behavior of the context assembly pipeline. Edit it on the wiki to change agent behavior without touching PHP files.

Key settings:

Setting	Default	Description
max_tokens	20000	Total token budget for assembled context
max_depth	2	Instruction dependency graph traversal depth
fully_resolve	false	If true, resolve entire dependency graph regardless of depth
max_related_pages	5	Pages from 1-hop link expansion (top-k by co-occurrence)
max_sources	5	Source documents retrieved per named entity
hop_depth	1	Link expansion hops (1 = direct links only)
include_talk	true	Include filtered talk page sections in context
use_wikipedia	false	Fetch Wikipedia intros for named entities
max_wikipedia_lookups	3	Maximum Wikipedia lookups per context build

Changes take effect immediately on the next agent call — no server restart required. If this page is missing or contains invalid JSON, all settings fall back to PHP defaults.